Inhibitory plasticity supports replay generalization in the hippocampus

Abstract

Memory consolidation assimilates recent experiences into long-term memory. This process requires the replay of learned sequences, although the content of these sequences remains controversial. Recent work has shown that the statistics of replay deviate from those of experience: stimuli that are experientially salient may be either recruited or suppressed from sharp-wave ripples. In this study, we found that this phenomenon can be explained parsimoniously and biologically plausibly by a Hebbian spike-time-dependent plasticity rule at inhibitory synapses. Using models at three levels of abstraction-leaky integrate-and-fire, biophysically detailed and abstract binary-we show that this rule enables efficient generalization, and we make specific predictions about the consequences of intact and perturbed inhibitory dynamics for network dynamics and cognition. Finally, we use optogenetics to artificially implant non-generalizable representations into the network in awake behaving mice, and we find that these representations also accumulate inhibition during sharp-wave ripples, experimentally validating a major prediction of our model. Our work outlines a potential direct link between the synaptic and cognitive levels of memory consolidation, with implications for both normal learning and neurological disease.

Figure 1.. Suppression of distractor stimuli in a random network model of CA3 replay.

a. Experimental paradigm and results adapted from Terada et al. (2022) (reprinted with permission). Calcium activity of CA3 Schaffer collateral axons was imaged with simultaneous contralateral local field potential (LFP) recordings to detect sharp-wave ripples (SWRs, upper left). Random cue-spatial task (RC-S): animals ran along a circular spatial belt, while three distractor sensory cues (odor, visual, and reward) were presented at pseudorandom locations to the mouse on every lap (upper right). CA3 units robustly represent both space (bottom left) and random cues (bottom center) in the online period. Bottom right: prior to the experiment, space and cue cells are indistinguishable in their SWR recruitment. In immobility epochs during the experiment, cue cells are suppressed from SWRs while place cells are recruited, an effect which persists in an immobility recording after the experiment. b. Network architecture during learning. Network consists of 8000 PCs (triangles) and 150 INs (circles) randomly recurrently connected. PCs and INs receive random cue and place input. Dashed lines indicate plastic synapses. c. Network during replay. Network now only receives random low-level spiking input (green arrows). d. Symmetric STDP kernel used for training the network: weight change $\text{Δ}W$ as a function of temporal difference $\text{Δ}t$ between pre- and postsynaptic spikes e. Learned synaptic weight matrices from a random sample of 400 place cells (left, sorted by location of peak tuning) and 50 cue cells (right, arbitrarily sorted), log scale. Place cells form strong weights to other place cells with peaks shortly before and after their own peak and weak weights with all other cells, while cue cells form strong weights with each other. f. Simulation of task in 1a (last 50 seconds of online learning and 10 seconds offline, slice of 10% of the PC network). Top: Simulated position and distractor inputs. Bottom: Raster of network activity during learning, sorted by cells’ tuning curve peak. Cue-responsive cells highlighted in yellow. Right inset: Spontaneous replay emerges in offline period. g. Neurons tuned to stochastic distractors (bottom) are active offline, but suppressed during replay.

Figure 2.. Distractor suppression in a biophysically detailed model of CA3 replay.

a. Schematic of the biophysical CA3 model. The model consists of excitatory PCs and perisomatic-targeting INs wired in accordance with experimental data. During online learning, the model is driven by multiple sources of structured external input – mossy fibers: place input; medial entorhinal cortex (MEC): grid input; lateral entorhinal cortex (LEC): cue input. During offline replay, the CA3 circuit is driven by random noise. SLM = stratum lacunosum moleculare; SR = stratum radiatum; SL = stratum lucidum; SP = stratum pyramidale; SO = stratum oriens. b. Spike raster of place cell (blue) and cue cell (orange) activity during online training during which STDP on $E\to E$ and $I\to E$ synapses is active. (left) At the start of training, the place cell ensemble poorly tiles the track and exhibits significant out of field firing. (right) After multiple laps, place cells effectively tile the linear track and exhibit little to no out of field firing. Orange arrowheads represent instances where cues input is provided to the network. c. Place cells are spatially tuned to distinct regions of the linear track after online learning. d. Somatic voltage potential for two example place cells showing in-field firing and ramp-like membrane potential depolarization. e. (top) Spike raster of place cells (blue, organized based on peak firing location on linear track) and cue cells (orange, no specific organization) during a period of high place cell activity show place cells with adjacent tuning peaks firing in rapid succession, consistent with a memory replay event. f. Somatic membrane potential of model place cells (bottom, in blue) and cue cells (top, in orange) during the period shown in e. illustrates that cue cells are suppressed when place cell activity increases. g. 2-D histogram of the number of spikes of cue cells and place cells are effectively decoupled during periods of high firing rate of place cells in the offline state. Two-sided Pearson’s $R=-0.19$, $p=4.36\times {10}^{-12}$, $n=1288$ time bins. h. Peri-event time histogram reveals that cue cell activity is suppressed across multiple high-synchrony events during the offline state. The orange line represents the mean cue cell activity, and the blue line represents the mean place cell activity. Shaded orange and blue areas represent respectively one standard deviation of cue cell and place cell activity. Two-sided Pearson’s $R=-0.23$, $p=2.12\times {10}^{-15}$, $n=1170$ time bins.

Figure 3.. Network perturbations and predictions.

a. Inhibitory weight onto cue cells is higher than place cells (median $w_{I\to cue}$: 0.0063 uS; median $w_{I\to place}$: 0.0046 uS; T = −26.37, $p=2.5\times {10}^{-187}$, one-sided Mann-Whitney U-test). b. Simulated inhibitory post-synaptic currents (IPSCs) recorded from place cells and cue cells in voltage clamp mode after a single interneuron spiked. Points represent mean IPSC magnitude, error bars represent one standard deviation of IPSC magnitude. Across a range of holding voltages, IPSCs were significantly higher in cue cells ($n=960$ pairs) than place cells ($n=4240$ pairs). One-sided Mann-Whitney U-test, ****p < 0.0001. At holding voltage −65 mV, $p=2.166008\times {10}^{-110}$; at −60 mV, $p=2.1660152\times {10}^{-110}$; at −55 mV, $p=2.1660152\times {10}^{-110}$; at −50 mV, $p=2.166024\times {10}^{-110}$; at −45 mV, $p=2.166036\times {10}^{-110}$. c. Feedforward inhibitory motif: cue cells are more suppressed by interneurons than place cells. d. Ablation of inhibitory plasticity: only $E\to E$ synapses are modified e. Average peri-event spike histogram of place cells (blue) and cue cells (orange) during spontaneous memory replay events showing that cue cell suppression is lost. Shaded orange and blue areas represent respectively one standard deviation of cue cell and place cell activity. Two-sided Pearson’s $r=0.002$, $p=0.96$, $n=780$ time bins. f. Heatmap of place-cue cell coupling: when inhibitory plasticity is ablated, the cue cell and the place cell subnetworks are coupled during the offline epoch (two-sided Pearson’s $r=0.1$, $p=0.71$, $n=1288$ time bins), in contrast to place-cue cell decoupling when inhibitory plasticity was active (Fig 2f, 2h). g. Schematic showing the sSTDP rule applied to $E\to E$ and $E\to I$ connections. h. Average peri-event time histogram of place cells (blue) and cue cells (orange) during spontaneous memory replay event showing that cue cell and place cell networks are tightly coupled during HSE. Shaded orange and blue areas represent respectively one standard deviation of cue cell and place cell activity. Two-sided Pearson’s $r=0.65$, $p=5.1\times {10}^{-60}$, $n=494$ time bins. i. Heatmap of place-cue cell coupling: when $E\to E$ and $E\to I$ plasticity is active, the cue cell and place cell subnetworks are tightly coupled during the offline epoch. Two-sided Pearson’s $R=0.75$, $p=6.25\times {10}^{-234}$, $n=1288$ time bins. Hence, $E\to I$ plasticity is insufficient for cue cell suppression.

Figure 4.. Abstract schematic of generalizable learning in the presence of noise.

a. Ground truth consists of an arbitrary sequence of stimuli to be learned. The “network” consists of “cells” each tuned to exactly one stimulus. In the ideal scenario, observations only consist of subsequences of the ground truth sequence. Symmetric STDP bidirectionally potentiates weights along the diagonal of the sorted weight matrix between cells representing adjacent elements of the sequence (right; cf. Fig 1e, left). b. A more realistic observation model contains noise: distractor stimuli may appear stochastically at different locations, or sequence stimuli may be erased or appear out of order. As a result, sSTDP potentiates off-diagonal weights (right; cf. Fig 1e, right). Under the realistic observation model of b., many off-diagonal synapses are potentiated. Thus, stimulating the first cell in the sequence offline will set off a stochastic cascade, possibly activating side paths through the network via distractor cells. Because distractors appear stochastically, they have positive weight on and may reactivate many other cells, poisoning structured replay. c. Active inhibition of these distractor cells reduces distractor cell firing rate (orange rasters) and rescues replay. As long as distractors are prevented from exciting other cells, the side paths are interdicted while the true sequence undergoes unencumbered reactivation, restoring structured replay. d. During each lap along the linear track, a distractor input is driven at a random location (tan boxes). This excitatory input directly targets PCs and INs within the local CA3 circuit, therefore increasing the firing rate for both cell populations for the duration that the distractor input is active. e. As a consequence of the distractor input increasing firing rates of PCs and INs, it is more likely that the delay between spikes from the PCs and the INs will decrease (i.e., reduction in $\left|pre\left(t\right)-post\left(t\right)\right|$). During this time, the sSTDP rule will generate a larger increase in inhibitory weight from INs onto PCs compared to epochs when the distractor input is turned off. f. For each lap, there will be a baseline increase in net inhibition onto place cells and cue cells $\left(I_{baseline}\right)$. Additionally, cue cells and place cells that were active during the distractor input will receive enhanced inhibition $\left(I_{enhanced}\right)$ for a net inhibitory change of $I_{baseline}+I_{enhanced}$. As the same population of cue cells are active at the time the distractor input is presented on each lap, the cue cell population will ‘accumulate’ this baseline and enhanced inhibition over $N$ laps $\left(I_{net,cue}=N{I}_{baseline}+N{I}_{enhanced}\right)$. In contrast, because the distractor input occurs at a random location on each lap, the enhanced inhibition will be distributed over the place cells, such that each place cell subpopulation will receive enhanced inhibition once over the course of $N$ laps $\left(I_{net,place}=N{I}_{baseline}+I_{enhanced}\right)$.

Figure 5.. Experimental verification of model predictions.

a. Viral strategy for sparse Cre-recombinase-driven expression of excitatory opsin (ChRmine) and dense expression of GCaMP in CA3/CA2 or in CA1. b. Example traces of stimulated (red: STIM) and unstimulated cells (black). Optogenetics stimulation (LED) evokes large-amplitude GCaMP-calcium responses (ΔF/F) in ins a subset of pyramidal cells (PCs) that express ChRmine (STIM-PCs), while other opsin-negative (Unstimulated) PCs show no response. c. Schematic of behavioral task. Optogenetic stimulation was triggered at odor cue presentations for the first 5 trials (Cue-STIM). The odor cue is randomly presented with 30 60 seconds interval (the same paradigm as in Terada et al. (2022)). As control, optogenetic stimulations (5x) were delivered randomly without odor cue presentation (Random). d. Left, Optogenetic co-stimulations with odor cue presentations induce cue-selective responses of both CA1 and CA2/3 PCs. Right, average responses of STIM-PCs (solid lines, Blue: CA1, Yellow: CA2/3) and unstimulated PCs (dash lines). e. Left, peri-SWR fluorescence (z-scored) of STIM-PCs before (Pre-Rest) and after (Post-Rest) optogenetic stimulation paired with odor cue presentation (top, Cue-STIM), and without odor cue presentation (bottom, Random). Right, averaged waveforms of peri-SWR fluorescence according to SWR component coefficient. f. Comparison of SWR component coefficients between Pre- and Post-Rest in Cue and Random conditions. PRE versus POST for CA1 Cue-STIM, n = 33, p = 0.00098, CA1 Random, n = 41, p = 0.000014, CA2/3 Cue-STIM, n = 35, p = 0.029, CA2/3 Random, n = 32, p = 0.00027 (Wilcoxon signed sum tests, n = 2 mice per group). Principal components of SWR responses are identified in CA1 and CA2/3 (see also Fig S8d). White line: median; box edges: 25th and 75th percentiles. Whiskers extend to the maxima and minima. Black lines: individual cells.